From variants to evidence

Altar represents three distinct kinds of scientific information:

1. Model scores are computed for the variants in an analysis.
2. Annotations are pre-existing per-variant facts or predictions retrieved from a reference source.
3. Variant–gene links are relations with their own locus, gene, context, score, and provenance.

Keeping these categories distinct answers an important question: did this analysis execute a model, look up a published value, or retrieve a regulatory relation?

Scientific workflow

input variants
    │
    ├─ validate against a reference genome
    │
    ├─ score with configured model instances
    │      └─ retain model identity and score schema
    │
    ├─ retrieve reference annotations
    │      └─ retain source version and missingness
    │
    ├─ retrieve variant–gene relations
    │      └─ retain biosample, method, distances, and provenance
    │
    └─ assemble per-variant results and evaluate declared rules

Altar does not collapse these inputs into one universal biological score. Downstream applications may rank, filter, or learn from them, but that policy must remain explicit and versioned.

What is shared across integrations

• canonical variant identity and occurrence handling;
• declared score and annotation schemas;
• model/source attribution;
• missing-value semantics;
• portable rule evaluation;
• execution and storage interfaces;
• conformance tests for extensions.

What is not shared is equally important: model-specific units, biological context, supported variant classes, and interpretation limits remain owned by each integration.